【作者】 赵彦玲；

【导师】 刘献礼；

【作者基本信息】 哈尔滨理工大学 ， 机械制造及其自动化， 2008， 博士

【摘要】 钢球作为轴承的主要零件,其表面缺陷对轴承的精度、运转性能和使用寿命等都有着至关重要的影响。目前,在轴承行业中,对钢球表面缺陷的检测仍然采用传统的人工目视检验法,其准确性和稳定性均难以保证。本文应用数字图像技术进行了钢球表面缺陷实时检测识别系统的基础性研究,给出了钢球缺陷识别有效性的评价体系。钢球表面有5种类型的缺陷,本文采用了高斯滤波和中值滤波分别对不同缺陷进行去噪,采用了灰度变换方法提高钢球图像边界信息对比度。根据钢球图像的特点,本文采用双阈值对缺陷进行分割,并提出了基于转基因遗传算子的OTSU理论自动优选阈值,清晰地将缺陷与背景分割开,完成图像二值化分割,达到计算机处理缺陷要求。由于分割后的钢球缺陷图像,存在缺陷边缘的不连续以及空间到图像映射不一致问题,导致缺陷边缘难以确定,为此本文提出了在小波域内利用小波变换多尺度分析和模局部极大值确定图像的边缘点并进行边缘点连接。实验结果表明,该方法检出的边缘完整、清晰,不需要进一步细化,同时避免噪声干扰。钢球在转动时其表面缺陷的面积是随其位置而变化的,因此仅靠平面图像中的面积进行缺陷分类会产生很大的误差,造成缺陷等级分类不明确及漏检现象。为消除这一影响,本文在大量实验的基础上,建立了钢球球体面积投射的校正模型,恢复了缺陷面积为钢球表面对应的实际面积。本文描述了钢球缺陷的形状特征参数:面积、长/短径比和欧拉数,给出了这些参数的计算方法,并对缺陷分类值进行定量分析。在对RBF神经网络训练算法深入研究的基础上,本文设计了基于RBF神经网络分类器对钢球缺陷进行识别。采用了两阶段学习策略来加速学习收敛;提出了动静相结合的隐含层设计算法构造较优的RBF神经网络结构;提出了误差校正的方法提高了RBF网络输出精度。开发了神经网络检测程序,并对点缺陷、凹坑缺陷、条缺陷和擦伤缺陷进行训练并测试,实验结果表明,基于RBF神经网络钢球缺陷识别法,准确率达96%。由于钢球是强反射球面,给获取清晰的缺陷图像带来困难,通过对钢球光学反射特性进行研究,结合大量的试验,设计了用柔光布制作的光照箱,完成了图像采集实验系统的搭建。针对钢球表面全检测的要求,本文设计了用于图像检测的钢球展开机构,对钢球在展开装置上的运动进行了分析,建立了拍摄点运动轨迹的数学模型,确定检测一粒钢球需要拍摄的次数。通过计算机仿真和实验验证了钢球在该机构上能够完全展开,从而达到了对钢球表面的全部检测。本文设计并开发了基于图像技术的钢球表面缺陷分析识别的软件系统,开发工具为VisualC++6.0,系统软件包括文件管理、图像处理算法、缺陷特征提取和缺陷判别4个功能模块,同时也可完成系统标定和分类器训练功能。更多还原

【Abstract】 Surface defects of steel balls, which are an important part of bearings, make a great difference to the accuracy, revolving performances and service life of bearings. Currently, it is a common practice to identify defects of steel balls by means of human visual checking whose accuracy and stabilization are hard to guarantee. This dissertation carries out a basic research on a real-time detection system of identifying steel ball surface defects with digital image technology, and provides an integral and effective evaluation system of identifying steel ball defects.There are five kinds of defects on steel ball surface, which appear at random. The dissertation first reduces noise through Gaussian filtering and median filtering, and then increases the contrast of edge information image through grey-scale transform. According to the property of steel ball images, this dissertation dismembers defects using dual threshold, presents automatically preferred threshold value based on a transgenic genetic GA genetic algorithms OTUS theory, thus clearly separating defects and background, completing image binarization separation, and enabling a computer to process testing. Because the edge of steel ball surface defects is discontinuous and inconsistent from space to plane, it is difficult to determine the edge of detects. This dissertation points out that multistage analysis and maximum of partial modulus can determine edge point of images based on wavelet transform and edge point link. Experimental results indicate that the method can get complete and clear edges and needs no further refinement, and at the same time avoid noise interference.The area of defect region varies with the location of defects while a ball is rotating, so it would result in great error, unclear classification of defects and omission of testing if defects were classified just according to areas of flat images. To remove this phenomenon, on the basis of a great number of experiments, this dissertation summarizes an emendating model of ball projected area, and resuming defect area is the responding surface area of a steel ball.This dissertation describes the shape characteristics of steel balls: area, ratio of long and short diameters, and Euler numbers, gives a calculation method, and quantitatively analyzes classified values of defects as well. Through further research on the training algorithms for RBF neural network, this dissertation designs a classifier based on RBF neural network to identify the defects of steel balls. Two-stage learning strategy is used to accelerate the rate of convergence, a preferable frame of implication layer by using a static and dynamic combination implication layer is put forward, and a algorithm is presented to improve the precision of RBF network output by means of error correction. Neural network programs are developed to train and test four kinds of defects: point defects, dent defects, strip defects and abrasion defects. Experiments have shown that accuracy is 96% based on RBF.It is hard to get a clear image of steel ball defects due to the complexity of steel balls’mirror reflection, so an illuminating system is designed, whose lighting box is produced with soft lighting cloth to finish the building of an image collecting experimental system. This dissertation designs a spreading device to detect steel balls, analyzes the motion of steel balls on the device, builds a math model of shooting point-motion trails, and determines the shooting times of testing a ball. Computer simulation and experiments have shown that this device can test the whole of the surface of a steel ball.This dissertation designs and develops an experimenting testing system based on image technology using VisualC++ 6.0 to identify defects of steel balls. The system consists of four-function modules: file management, image processing algorithm, defect property extraction, and identification of defects.更多还原